Title: Probabilistic Upper Bounds for Urgent Applications Nick Trebon and Pete Beckman University of Chicag
1Probabilistic Upper Bounds for Urgent
ApplicationsNick Trebon and Pete
BeckmanUniversity of Chicago and Argonne
National Lab, USA
Case Study
Elevated Priority Experiments
Normal Priority Experiment
Conclusions
In order to evaluate the individual phase and
composite bounding methodologies, a case study
was performed using the FLASH scientific
application on the TeraGrid. While FLASH is not
a typical urgent computation due to the lack of a
deadline constraint, it is a complex and
well-known scientific code.
The first experiment examined the performance of
the bounds for a normal policy (i.e., no SPRUCE).
The next two experiments examined the performance
of the bounds for the next-to-run and preemption
policies. Only the UC/ANL resource was part of
this experiment. The 0.95 quantile was targeted
for each individual phase, resulting in an 0.815
composite target quantile.
- Preliminary approach to generating
empirical-based upper bounds on total turnaround
times for urgent applications performs well for
elevated priority experiments. - The overprediction in the normal priority queue
phase is most likely caused by skew in batch
history. SPRUCE elevated priorities provide
users with individual and composite bounds that
are both accurate and correct. - The composite bounds can be used to guide urgent
computing users in selecting a resource with
greater confidence that their deadlines will be
met.
Performance of Composite Bounds for Normal
Priority (Target Quantile 0.815)
Performance of Composite Bounds for Next-to-Run
and Preemption Policies (Target Quantile 0.815)
Case Study Computational Resources
Note difference in scales
Each individual phase targeted the 0.95 quantile.
Thus, the composite quantile was 0.815.
Because Flash does not have any file staging
requirements, these were artificially added. The
requirements were loosely modeled after Linked
Environment for Atmospheric Discovery (LEAD)
workflow. LEAD is a project that, in 2007,
teamed with SPRUCE in order to perform real-time,
on-demand, dynamically adaptive forecasts.
Success Rate for Individual and Composite Phases
Success Rate for Individual and Composite Phases
for Elevated Priorities
References
- N. Trebon, Deadline-based grid resource
selection for urgent computing, Masters thesis,
University of Chicago, Chicago, IL Jun. 2008. - SPRUCE http//www.spruce.uchicago.edu
Overprediction Rate for Individual and Composite
Phases
Overprediction Rate for Individual and Composite
Phases for Elevated Priorities
Case Study File Staging Requirements